GAME E-COMMERCE WEBSITE WITH PRODUCT RECOMMENDATION SYSTEM AND AI CHATBOT

Abstract

The rapid expansion of e-commerce platforms has intensified the problem of information overload, making it increasingly difficult for users to discover products aligned with their preferences. Conventional search and recommendation methods often lack semantic understanding of user intent and product attributes, leading to reduced recommendation accuracy and user satisfaction. This thesis presents the design, implementation, and evaluation of an intelligent e-commerce platform that integrates modern artificial intelligence techniques to deliver personalized recommendations and conversational shopping support. The proposed system is implemented as a scalable three-tier web architecture comprising a React and TypeScript frontend, a Node.js/Express backend providing RESTful services, and a PostgreSQL database enhanced with the pgvector extension for vector similarity search. This architecture enables efficient management of users, products, interactions, and transactions while maintaining responsive system performance. The core contribution of this research is a hybrid recommendation engine that combines content-based and collaborative filtering approaches. Content-based recommendations utilize 768-dimensional semantic embeddings generated by Google’s text-embedding-004 model to capture product meaning and user preferences, while collaborative filtering exploits implicit behavioral signals such as views, purchases, and reviews. The two methods are fused using weighted score normalization, allowing the system to balance personalization accuracy and recommendation diversity. Vector similarity search is accelerated using HNSW indexing, achieving low-latency recommendation retrieval on large product catalogs. To further enhance usability, the platform integrates a conversational AI chatbot based on a Retrieval-Augmented Generation (RAG) architecture. The chatbot supports natural language product search, recommendations, and customer support through function calling and semantic caching, significantly reducing response latency for repeated or similar queries. Comprehensive evaluation results demonstrate strong system performance, with API response times under 100 ms, recommendation generation within 200 ms, and high recommendation quality metrics, including Precision@10 of 0.62 and NDCG@10 of 0.71. User testing confirms positive satisfaction with both usability and recommendation relevance. Overall, the research validates the effectiveness of hybrid recommendation techniques combined with conversational AI in modern ecommerce systems.